[Xenomai] select() unblocks when no data to read()

* [Xenomai] select() unblocks when no data to read()
@ 2017-07-13 16:26 George Broz
  2017-07-13 20:39 ` George Broz
  0 siblings, 1 reply; 4+ messages in thread
From: George Broz @ 2017-07-13 16:26 UTC (permalink / raw)
  To: xenomai

Hi -

I'm using the UDD driver on Xenomai 3.0.4 / Linux 3.18.20 on an ARM SoC.

I use select() to block on read events. The problem is select() unblocks
twice in response to one interrupt event. The corresponding read()
to the second erroneous unblocked select() then blocks until the next
interrupt arrives.

The psuedo-code in user-space looks like this:

  while (active)
  {
    FD_ZERO (&FileDescSet);
    FD_SET (FileDesc, &FileDescSet);

    /* block until interrupt or timeout occurs */
    IsInterruptPending = select ((FileDesc + 1), &FileDescSet, NULL,
NULL, &Timeout);

    /* check timeout (0) and error cases (-1) */
    ...

    /* otherwise, process */
    if (IsInterruptPending && FD_ISSET(FileDesc, &FileDescSet))
    {
      err = read (FileDesc, &u32EventCount, sizeof (u32EventCount));
      ...
    }
  }

I have added debug to the UDD driver to verify udd_notify_event() is called
exactly once per interrupt event, and that the ur->event is properly
incremented.

In udd_read_rt(), debug showing ur->event and context->event_count do what
they should be doing - returning data when they mismatch and then setting
one equal to the other, or blocking when they are equal. This function would
be fine if select were working properly.

(One thing to note is that when ur->event and context->event_count are equal,
it takes two iterations of the for (;;) loop to actually block on
rtdm_event_wait().)

I added debug to udd_select(), but it doesn't work as expected. I had hoped
to see debug emitted for each user-space call to select(), but it seems only
one call is made to udd_select() the first time through the loop.

In an unsuccessful workaround, I tried issuing rtdm_event_init() in an
ioctl() called just before unmasking interrupts for the next iteration
of the loop.

Is there a possible bug in the RTDM select() handling? Is there a patch
or additional debug someone could suggest? Any input greatly appreciated.

Regards,

--George Broz
Moog Industrial Group

^ permalink raw reply	[flat|nested] 4+ messages in thread